합성곱 신경망과 머신러닝을 활용한 화물운반 주의 표식 인식

Abstract

The development speed of the loading and unloading automation is accelerating due to such initiators as the prevailing 4th industrial revolution technologies, the expansion of e-commerce, and the increase in logistics volumes inevitably by the advent of the corona era. The logistic-object classification evolves along with the computer vision as a core technology for the automatic loading and unloading. As a practical example, ‘Empticon Ⅱ’ of the ‘Roblog’ project identifies the location of the coffee bean sack to be transported. The 3D camera is used for photographing the sack and analyzing it through the computer vision technology. Tremendous amounts of data training efforts were required in advance to make the object detection function better. This method, however, has a disadvantage in that it is difficult to learn and utilize all the objects of various sizes and textures in advance because the object detection performance is proportional to the amount of objects learned in advance. Most of the objects handled are boxes. To prevent damage from using the unloading robot, logistics safety labels are attached to the box. This study aims to make a model that can control automatic loading and unloading robots by recognizing and classifying logistics safety labels attached to boxes of various sizes. To train the machine learning model, the most common logistics safety labels attached to the box or printed need to be obtained as training examples. Unfortunately, there was no such dataset being distributed as an open source, so the authors built a dataset by themselves. A classifier using the ensemble model was created. In order to increase the efficiency of the classifier, the characteristics of the logistics safety labels were extracted through a convolutional network and used as training and test data. As a result, we observed the performance of the ensemble classifier having Precision 0.9675, Recall 0.9575, and F1-score 0.9625. In addition, in order to solve the problem of performance degradation of the classifier depending on the background in which the box is photographed, a pipeline was designed to detect only the box without background and caution signs.